fix(auto labeling 3d) webauto integration fixes

[jacoblambert]

Summary

PR to fix various bugs and update configs in auto_labeling_3d in the process of merging with webauto.

major bug fix: remove temporally aligned frames assumptions

• Previous implementation assumed each model produced output that was temporally aligned.
• This is not true for camera based models since cameras are often dropped.
• Now we align frames by token and treat a missing model/frame pair as an empty prediction set instead of accidentally shifting later frames into the wrong lidar timestamp

major bug fix: add yaw in NMS calculation:

• NMS calculation was previously done without considering yaw, just axis aligned.
• This is fast but approximation that lead to many duplicate boxes
• 2D rotated bbox NMS is not very slow and there is no real performance constraint, so I re-implement it properly
• I also adjust ensemble parameters to suppress overlapping boxes more aggressively.

minor bug fix: do MOT without time truncation

• t4-devkit enforces int unix time, which was then used for tracking
• work around this by storing the float timestamp in the tracker
• probably very minor but no real reason to introduce this error

chore: update auto_labeling_3d entrypoint with clean config.py for each model.

• Uploaded auto_labeling_3d configs to model_zoo
• Updated file paths to avoid too many scattered tmp files
• Updated score filtering for latest model
• Added missing polars dependency

[tughril]

LGTM

[KSeangTan]

LGTM overall, please resolve the comments and merge it afterwards